Abstract
Facing the real-life needs for personalized interaction capable of assisting learners in actively engaging with a variety of language resources to promote their own development, language educators have been expecting Intelligent Language Tutoring Systems (ILTS) as the ultimate solutions that capture the interdisciplinary quintessence of Human Language Technology. However, the majority of current ILTS do not address the innovative changes in language teaching methodology that supports the balance between the structural knowledge of language (form-focused) and the ability to use it for practical communication purposes (meaning-focused). Specifically, they tend to heavily focus on form-based interaction such as corrective feedback on grammatical errors. In this context, we aim to develop a text-based dialog system that is capable of acquiring semantic and pragmatic knowledge to create meaning-focused interaction for English language learners (ELLs).\r Specifically, we first develop a robust semantic representation of a online authentic text that is integrated with plenteous meaningful features for language learning purposes and then design learning units as coherent sequences of individualized meaning-focused interactions between a conversational agent and a learner, based on the developed semantic representation of the core text. To address the first task, we chose the sentence-level Abstract Meaning Representation (AMR) formalism as the starting point and make an pioneer attempt at solving the problem of discourse level coreference resolution for AMR texts as sequences of AMR graphs. Inspired by Centering theory and Stanford’s rule-based coreference resolution system, the proposed algorithm features a simple cache model for searching antecedent candidates, a versatile and scalable framework for coreference feature integration, and a multiple-factor model of salience for ranking antecedent candidates. The output cross-sentential co-referred concepts are then merged together to develop the document-level AMR graph for the examined text. Based on this robust text-based semantic representation, we implement a proof-of-concept dialog system handling Semaland game, an exclusive in-house design that allows ELLs to truly interact with the system’s conversational agent regarding the semantic concepts in the text they have read, and therefore provides the learners with beneficial individualized learning experiences.